A system that doesn't thake decision is useless

May 2025

just because of one incident, I had a swing of false realization that rule-based thresholds are the most useless to the most useful thing in the world. Well it's neither of them. Let me explain...

Context

I've deployed the initial version of Layer 1 (L1) with Dimensions and it was just... useless. It gives you random numbers that don't mean anything. Even as the guy who built this, I didn't understand what the hell was going on.

What Happened?

the whole idea is that I have data and to find data issues, I use signals. These signals extract info from the data and that info is processed and then a decision is made from the final result.
the problem is on the decision part. I was trying to design a system that doesn't make any wrong decisions. So in order to achieve this impossible goal, I started to build workarounds and add complexities. In the first version, in order to achieve this goal, I somehow accidentally removed the decision-making in the first place.
the simplest way the system can make a decision is using thresholds (if-else). Just define the condition and what decision to take for that. This is fine to start with, but as my instinct goes, I hated it. I wanted something god-like, ultra-complex, a freaking brain-type decision-making that will replace the human... (kinda).
so I switched to a score-based system. Now I don't need to define a condition, I can just define the severity of the problem. Much like a neuron in a neural network — weights × inputs.
this worked, I completed L1, shipped. Realized it's shit. It was shit because there are numbers floating on my screen like traffic lights and telling me this is a problem because it's > 0.78. The system was not making any decision at all.
so I redesigned the entire thing making sure that now it makes decisions and takes a stance. For that I tried the simplest approach (crude if-else thresholds) to understand why it doesn't work and to break my bad habit of adding abstraction.
it worked surprisingly well. I didn't understand why, but using universally accepted thresholds and high-quality signals were enough to even make a simple workflow like this perform great.

What I Found

high-quality signals + universally accepted thresholds are the reason why crude thresholds worked for L1.
this won't be the case for L2 or other downstream layers because there will be noisy, overlapping, and weak signals.
a system that doesn't make a decision is useless. Especially an ML system.
taking a wrong decision is better than taking no decision and hiding behind ambiguity. In most cases, wrong decisions can be made correct.

Constraints

these are still just basic if-else conditions that are tested on a lot of data and real-world cases to make sure they are robust.
these are basically drawing hard decision boundaries. These hard decision boundaries work here but not everywhere.
these hard boundaries can make wrong decisions.
there is no learning here.